Product / Monitor
Every run, step by step.
Each run keeps a trace: what it read, what it sent, how long it took and what it cost. When something fails, it retries once, then hands the run to its owner with the trace attached.
Weekly churn digest
Last 4 runs
| Run | Started | Result | Time | Cost |
|---|---|---|---|---|
| #418 | Mon 09:00 | Done | 41 s | $0.011 |
| #417 | Mon 09:00 | Done | 38 s | $0.010 |
| #416 | Mon 09:00 | Retried once | 1 m 12 s | $0.019 |
| #415 | Mon 09:00 | Done | 44 s | $0.012 |
What Monitor does
01
Traces you can follow
Open any run and read it top to bottom: every tool call with its input and output, every decision with its reason.
kept 7 to 400 days by plan
02
Cost per run
Model tokens and tool calls priced on every step, so you know which agent costs what, and why.
$0.011 a run on average
03
Alerts to a person
Failures retry once. Then the owner hears about it in Slack or email, with the trace and what was not done.
Slack · email · webhook
Read a run like a log.
Timestamps, tool calls, model use and the approval, in order. Search across runs by agent, tool, customer or error, and export to your warehouse.
- 09:00:01run #418 started · schedule
- 09:00:02stripe.subscriptions.list → 214 rows
- 09:00:06hubspot.companies.search → 31 of 31
- 09:00:14model · 2,140 tokens · $0.004
- 09:00:15waiting · approval from Maya Okafor
- 09:02:41slack.chat.post #revenue → ok
Failures go to a person, not a void.
When QuickBooks rate-limits or a tool times out, Capstan retries once, then stops and tells the agent's owner what happened and what it didn't do.
- QuickBooks answered 429: too many requestsquickbooks.invoices.query · 08:00:04
- Retried after 60 s, same answerretry 1 of 1 · 08:01:05
- Told Lena Hoffmann in Slack, with the tracenothing was sent to customers